High pooled performance of convolutional neural networks in computer-aided diagnosis of GI ulcers and/or hemorrhage on wireless capsule endoscopy images: a systematic review and meta-analysis

Background and Aims: Diagnosis of GI ulcers and/or hemorrhage by wireless capsule endoscopy (WCE) is limited by the physician-dependent, tedious, time-consuming process of image and/ or video classification. Computer-aided diagnosis (CAD) by convolutional neural network (CNN)-based machine learning may help reduce this burden. Our aim was to conduct a meta-analysis and appraise the reported data. Methods: Multiple databases were searched (from inception to November 2019), and studies that reported on the performance of CNN in the diagnosis of GI ulcerations and/or hemorrhage on WCE were selected. A random-effects model was used to calculate the pooled rates. In cases where multiple 2 × 2 contingency tables were provided for different thresholds, we assumed the data tables were independent from each other. Heterogeneity was assessed by I²% and 95% prediction intervals. Results: Nine studies were included in our final analysis that evaluated the performance of CNN-based CAD of GI ulcers and/or hemorrhage by WCE. The pooled accuracy was 95.4% (95% confidence interval [CI], 94.3-96.3), sensitivity was 95.5% (95% CI, 94-96.5), specificity was 95.8% (95% CI, 94.7-96.6), positive predictive value was 95.8% (95% CI, 90.5-98.2), and negative predictive value was 96.8% (95% CI, 94.9-98.1). I²% heterogeneity was negligible except for the pooled positive predictive value. Conclusions: Based on our meta-analysis, CNN-based CAD of GI ulcerations and/or hemorrhage on WCE achieves a high-level performance. The quality of the evidence is robust, and therefore CNN-based CAD has the potential to become the first choice of machine learning to optimize WCE image/video reading.